Research map / Label-efficient and robust learning / Recognition and long-tail
Online Boosting Adaptive Learning under Concept Drift for Multistream Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OBAL, a dual-phase online boosting ensemble for multistream classification that uses AdaCOSA (adaptive re-weighted CORAL-style covariate alignment) for initialization and GMM-based weighting plus DDM and two-window drift detection for online asynchronous drift adaptation.
Problem
Multistream classification with unlabeled target stream: covariate shift, asynchronous concept drift, scarce target labels, and temporally varying correlations between source and target streams causing negative transfer.
Approach
Adaptive re-weighting of archived source instances combined with CORAL-style covariance alignment (AdaCOSA) learns time-varying source-target correlation weights from target-domain predictions; an online GMM-based weighting and classifier pool handle asynchronous drift, with re-initialization on detected target drift.
Key results
- OBAL reaches 90.98% accuracy on SEA and 86.78% on RBF, outperforming the best baseline (AOMSDA 90.23% on SEA; Melanie 86.04% on RBF).
- OBAL achieves 70.29% on Kitti versus 67.79% for AOMSDA, and 70.71% on CNNIBN versus 69.07% for AOMSDA.
- Ablation shows replacing no covariate alignment with CORAL and then with AdaCOSA improves accuracy on every dataset (e.g., SEA 82.48 -> 88.76 -> 90.98).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OBAL instead of the strongest listed baseline | multistream classifiers · SEA | accuracy (%): 90.23 -> 90.98 | Helped |
| use OBAL instead of the strongest listed baseline | multistream classifiers · Tree | accuracy (%): 78.93 -> 78.45 | Hurt |
| use OBAL instead of the strongest listed baseline | multistream classifiers · RBF | accuracy (%): 86.26 -> 86.78 | Helped |
| use OBAL instead of the strongest listed baseline | multistream classifiers · Hyperplane | accuracy (%): 87.66 -> 88.01 | Helped |
| use OBAL instead of the strongest listed baseline | multistream classifiers · Weather | accuracy (%): 77.74 -> 79.22 | Helped |
| use OBAL instead of the strongest listed baseline | multistream classifiers · Kitti | accuracy (%): 67.79 -> 70.29 | Helped |
| use OBAL instead of the strongest listed baseline | multistream classifiers · CNNIBN | accuracy (%): 69.07 -> 70.71 | Helped |
| use OBAL instead of the strongest listed baseline | multistream classifiers · BBC | accuracy (%): 68.04 -> 66.43 | Hurt |
| add drift and covariate-shift adaptation (ablation) | OBAL · all eight benchmarks | accuracy (%): OBAL_v1: 79.54, 72.42, 79.78, 81.42, 72.25, 62.14, 63.12, 58.02; OBAL: 90.98, 78.45, 86.78, 88.01, 79.22, 70.29, 70.71, | Helped |
| add covariate-shift alignment (ablation) | OBAL with synchronous drift adaptation · all eight benchmarks | accuracy (%): OBAL_v2: 82.48, 74.74, 81.41, 82.35, 74.18, 64.09, 67.44, 62.77; OBAL_v3: 88.76, 77.01, 84.23, 86.34, 77.43, 68.24, 69.0 | Helped |
| add adaptive correlation weighting instead of standard CORAL (ablation) | OBAL · all eight benchmarks | accuracy (%): OBAL_v3: 88.76, 77.01, 84.23, 86.34, 77.43, 68.24, 69.01, 64.12; OBAL: 90.98, 78.45, 86.78, 88.01, 79.22, 70.29, 70.71, | Helped |
Surprising findings
- Melanie, a supervised method using target labels, is beaten by OBAL on only six of eight datasets, and is stronger on Tree and BBC.
- Performance with five source streams is reported to exceed that with seven on the Tree dataset, suggesting more sources can hurt.
Limitations
- Relies on a Hoeffding Tree base classifier and a fixed set of hyperparameters (L_n, I_max, pool size) tuned per dataset.
- Target-drift detection triggers full re-initialization, discarding all base classifiers.
- Source-count experiments show possible performance decline with more streams, attributed to model complexity without deeper analysis.
Experimental setup
Datasets: SEA Tree RBF Hyperplane Weather Kitti CNNIBN BBC
Benchmarks: SEA Tree RBF Hyperplane Weather Kitti CNNIBN BBC
Metrics: classification accuracy execution time
Baselines: FUSION ATL Melanie AOMSDA MCMO
Builds on: CORAL FUSION Melanie AOMSDA MCMO
Closest papers by meaning
- Test-time Adaptation in Non-stationary Environments via Adaptive Representation Alignment (NeurIPS 2024 · similarity 0.84)
- AdaO2B: Adaptive Online to Batch Conversion for Out-of-Distribution Generalization (AAAI 2025 · similarity 0.84)
- Early Concept Drift Detection via Prediction Uncertainty (AAAI 2025 · similarity 0.83)
- Label Shift Meets Online Learning: Ensuring Consistent Adaptation with Universal Dynamic Regret (CVPR 2025 · similarity 0.83)
- Fast Co-Training under Weak Dependence via Stream-Based Active Learning (ICML 2024 · similarity 0.83)
- Stabilizing Linear Passive-Aggressive Online Learning with Weighted Reservoir Sampling (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.